Research And Application Of Fabric Defect Detection Algorithm

This paper focuses on the textiles with periodicity changing patterns,and studies the detection algorithm of the defective parts.In the field of traditional image processing,aiming at the stretching and rotation deformation of textile images in the process of shooting and generation,an image correction method is proposed,which corrects the irregular patterns with different degrees into neat patterns in accordance with the periodic law of the image itself.In the field of deep learning,SSIM loss function and Gaussian noise term are added on the basis of traditional autoencoders to improve the ability of image reconstruction,which can be used to reconstruct defective images into flawless ones to realize the recognition of defective parts.Using the ability of deep learning classification network for semantic segmentation of image content,this paper tries its application in textile defect detection and obtains good results.This paper mainly studies the image preprocessing,image reconstruction of autoencoder and semantic segmentation network in textile defect detection.The main research contents are as follows:(1)Study on image correction method.In view of the interference caused by different degrees of deformation in textile images to unit segmentation,an image correction method based on Hough transform and perspective transform is studied.The irregular patterns caused by different rotation,stretching and cutting methods are corrected to unit patterns in line with the image period.It is helpful for further image segmentation and template correction.At the same time,a more uniform shape standard of textile image can be obtained by correcting the original image.(2)Study on image segmentation and template correction methods.For textile images with periodic patterns,the unit patterns in shading have certain periodic rules in both horizontal and vertical directions.Therefore,the textile image is segmented according to the period,the pattern information is contained in the form of units,and the units are further corrected to achieve the purpose of decomposing the whole image into multiple identical cells.The feature extraction and defect recognition are carried out with cells as the unit,which greatly reduces the spatial complexity of the algorithm.Moreover,the similar correlation between cells can better distinguish the difference between the defective part and the shading background.(3)Research on image reconstruction of autoencoder.Through the analysis of the essential problem that the defective part of the defective image is the noise added on the shading background,considering the image reconstruction ability of the autoencoder,the defective image can be reconstructed into a flawless image.The advantages and disadvantages of several traditional autoencoders are analyzed,and the corresponding improvements are made on the basis of them,and the textile defect image is adjusted in order to adapt to the task of defect detection,and good experimental results are obtained.(4)Research on different semantic segmentation and classification networks.In view of the different kinds of shading and different types of defects in the textile defect images,this paper tries to use semantic segmentation network to classify the textile images,and at the same time to identify the defect parts for further accuracy measurement.The application of Deep Lab V3 +,GCN,PSPNet,Bi Se Net and ICNet methods to textile defect detection has been realized,and good experimental results have been achieved in the classification accuracy and the accuracy of defect identification.